Exploding hedging errors for digital options (Q1297920)

Attainable contingent claims in a frictionless market are priced by the following no-arbitrage argument: by the definition of attainability, there exists a self-financing strategy for trading a portfolio of the underlying risky asset and a risk-free bond such that the payoff of the contingent claim is exactly duplicated by the value of the portfolio. From this it is concluded that the market price of such contingent claim must equal the initial value of the self-financial portfolio trading strategy, since any other price would give a clever investor an obvious arbitrage opportunity. Trading in a frictionless market, that is continuously and without transaction costs, the arbitrage could buy or sell short the contingent claim and use the above self-financing trading strategy to hedge his position completely. This no-arbitrage argument, developed by \textit{F. Black} and \textit{M. Scholes} [J. Polit. Econ. 81, 637-659 (1973)], and self-financing strategy is often referred to as the delta hedging strategy for the contingent claim. This article argues that, even under the assumption of a frictionless market, the delta hedging strategies for certain contingent claims, like digital or barrier options, require the arbitrage to know the exact probability distribution of the underlying risky asset to avoid making a substantial loss under any circumstances. If the model of geometric Brownian motion is applied to calculate the delta hedging strategy for a European digital option, it is shown that there exists a continuous, arbitrage-free stochastic process for the underlying risky asset which causes the delta hedging strategy to produce disaster results: with probability arbitrarily close to one, the difference \(\Delta\) between the payoff of the contingent claim and value of the hedge portfolio at maturity becomes arbitrarily large. Furthermore, it is shown that even if the underlying risky asset follows a geometric Brownian motion but with a volatility different from that assumed in the model, then the variance of \(\Delta\) can become very large.